Same Pre-training Loss, Better Downstream: Implicit Bias Matters for Language Models.

A connected record in the developer community graph.